Overview
- OpenAI for Science published case studies on November 20 showing GPT-5 shortens parts of scientific workflows when used by experts, without running projects autonomously.
- The paper reports four new mathematics results credited to GPT-5’s assistance and verified by human authors, including a missing step that helped close an Erdős problem.
- In an immunology project at The Jackson Laboratory, GPT-5 identified a likely cause from unpublished data within minutes and proposed an experiment that the lab confirmed.
- Researchers describe large time savings for coding and literature review, with the model surfacing relevant work across decades and languages that humans had not found.
- OpenAI and outside scientists caution that GPT-5 can hallucinate citations or mechanisms, stress the need for expert oversight, and note the case‑study design limits reproducibility and does not indicate imminent AGI.